Attention-Driven Multisensor Fusion for Accurate and Real-Time Wind System Fault Diagnosis

This paper presents an advanced hybrid fault detection and diagnosis (FDD) framework for wind energy systems that integrates real sensor measurements with Digital Twin (DT) representations through hierarchical multi-sensor fusion, graph-based learning, and attention-driven ensemble modeling. The framework employs a Global Fusion (GF) mechanism to unify real and synthetic feature spaces, enhancing data diversity and enabling effective cross-domain knowledge transfer. A Multimodal Attention (MA) scheme further boosts discriminative power by combining deep latent embeddings with globally fused features, which are then processed through an optimized bagging ensemble to ensure robust generalization and computational efficiency. Extensive experiments on wind turbine datasets demonstrate the robustness and superiority of the proposed framework. The MA framework achieves perfect classification metrics across all evaluation criteria (accuracy, precision, recall, F1-score), outperforming traditional and modern approaches, including Random Forest (RF), k-Nearest Neighbors (kNN), Decision Tree (DT), Naive Bayes (NB), Support Vector Machine (SVM), and Graph Convolutional Network (GCN) models. These results underscore the effectiveness of attention-driven integration of sensor, DT, and relational features in delivering a scalable, high-performance, and interpretable FDD solution for contemporary wind energy systems.